Aplicação de modelo de Rede Neural Convolucional para detecção de fogo ativo no Parque Estadual Encontro das Águas

Abstract

The detection of active fire through remote sensing is an important tool for controlling and analyzing the impacts of forest fires, especially when they occur in remote and difficult-toaccess regions. With technological advancement, image processing algorithms emerged and increased the alternatives to monitor and analyze these disasters. Thus, the objective of this research is to apply a Convolutional Neural Network on images of a region of interest that includes the Encontro das Águas State Park from August to October 2020, a period in which the reserve – the second largest in the Brazilian Pantanal – suffered from a historic number of fires. The data used in the experiment was acquired from the Sentinel-2 satellite, through Python’s eo-learn library, resulting in a total of fifteen image, with spatial resolution of twenty meters and 2496 x 2489 pixels. At the end of the preprocessing, 1500 images of 256 x 256 pixels were obtained, which were passed through the chosen Convolutional Neural Network. The model used corresponded to a U-Net, which had nine convolution blocks, with 2,164,593 pre-trained parameters. After processing, fifteen masks of size 2560 x 2560 were obtained, each corresponding to a date in the period analyzed, whose pixels indicate the presence of active fire, when True, or not, when False. The analysis of the masks indicated that the peak of the wildfires in the region occurred in September, the last month of the dry period in the region. The four images acquired between 08/29/2020 and 09/13/2020 accumulated 84.7% of all active pixels in the period, with a peak on the last day, which had 20,177 activated pixels, equivalent to approximately one third of the total for the period. After that date, the fires detected by image had a sharp drop, which continued until the end of the analyzed period. To validate the results found, a comparison was made between the generated masks and the results obtained on the BDQueimadas platform, for the same dates and delimited area, which indicated a greater capacity for detecting smaller and more detailed fire spots by the studied network. Thus, the tool proved to be useful for further analysis of the evolution and impacts of fire in large areas. However, it was considered that despite Sentinel- 2's higher spatial resolution, its five-day temporal resolution, in addition to its processing time, make it unfeasible to use for real-time monitoring of these phenomena.